Approximate SRAM for Energy-Efficient, Privacy-Preserving Convolutional Neural Networks
Lita Yang, Boris Murmann · 2017
Convolutional neural networks (ConvNets) achieve remarkable accuracy in a variety of classification domains. Unfortunately, deployment of these algorithms in embedded systems and mobile devices is limited by the high memory power consumption caused by network storage requirements and memory access intensity. Our work studies the tradeoff between energy (required memory supply voltage), accuracy (classification error rate), and privacy (noise injection mechanisms) in hardware implementation of ConvNets. Our implementation and experiments based on measurements from a 28nm SRAM test chip demonstrate supply voltage reduction of 300mV at 99% of floating-point classification accuracy, under a modest (9, 10-5)-differential privacy budget. This corresponds to 5.0x leakage power reduction and 2.8x memory access power savings, with minimal hardware overhead.